Rethinking Model Calibration through Spectral Entropy Regularization in Medical Image Segmentation

Kun Cheng, Yukun Zhang, William Henry Nailon, Tonggang Zhao

Keywords: medical image segmentation, model calibration, spectral entropy

TL;DR: We improve medical image segmentation model calibration by addressing spectral bias and confidence saturation through a frequency-domain framework that balances low and high-frequency components in uncertainty estimation.

Abstract: Deep neural networks for medical image segmentation often produce overconfident predictions, posing clinical risks due to miscalibrated uncertainty estimates. In this work, we rethink model calibration from a frequency-domain perspective and identify two critical factors causing miscalibration: spectral bias, where models overemphasize low-frequency components, and confidence saturation, which suppresses overall power spectral density in confidence maps. To address these challenges, we propose a novel frequency-aware calibration framework integrating spectral entropy regularization and power spectral smoothing. The spectral entropy term promotes a balanced frequency spectrum and enhances overall spectral power, enabling better modeling of high-frequency boundary and low-frequency structural uncertainty. The smoothing module stabilizes frequency-wise statistics across training batches, reducing sample-specific fluctuations. Extensive experiments on six public medical imaging datasets and multiple segmentation architectures demonstrate that our approach consistently improves calibration metrics without sacrificing segmentation accuracy.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 1204